bev-decider

Run bev-decider-0.4B, a 0.4B-parameter System One decision model. It reads a state (text or JSON) and typed questions about it, and returns calibrated probabilities in a single forward pass. It uses TypeSafe Jev's question and answer format, so a local server can stand in for the /v1/systemone API.

• 0.4B parameters. It runs on a laptop CPU, Apple Silicon or any GPU.
• Choice-order invariant. Options are read in parallel from the same position, so reordering them cannot change the answer.
• Typed answers: choice (a key and probabilities), noul (P(yes)) and score (an expected level and probabilities).

See the model card for benchmarks and known weaknesses.

Install

pip install bev-decider            # library
pip install "bev-decider[serve]"   # + local /v1/systemone server

The weights (about 35 MB) and the base Qwen/Qwen3-0.6B are downloaded from the Hugging Face Hub on first use.

Python

from bev_decider import load

decider = load()  # downloads avbiswas/bev-decider-0.4B

state = {
    "message": (
        "URGENT: you charged my card twice this month. "
        "Refund the duplicate within 24 hours or I'm disputing it with my bank."
    )
}

questions = {
    "intent": {
        "type": "choice",
        "instructions": "What does the customer want?",
        "criteria": {
            "refund": "money returned or a duplicate charge reversed",
            "technical_help": "a bug, outage or integration problem",
            "cancellation": "wants to cancel or downgrade",
        },
    },
    "urgent": {
        "type": "noul",
        "instructions": "Does the message communicate time pressure or a deadline?",
    },
    "anger": {
        "type": "score",
        "instructions": "How angry is the customer?",
        "criteria": ["calm", "mildly annoyed", "frustrated", "furious"],
    },
}

answers = decider.decide(state, questions)

Output:

{
  "intent": {
    "type": "choice",
    "choice": "refund",
    "probabilities": {"refund": 1.0, "technical_help": 0.0, "cancellation": 0.0}
  },
  "urgent": {"type": "noul", "noul": 0.99},
  "anger": {
    "type": "score",
    "score": 1.88,
    "probabilities": {"0": 0.09, "1": 0.10, "2": 0.64, "3": 0.16}
  }
}

To choose a device, pass load(device="cpu"), "mps" or "cuda".

Question types:

Questions about the same state are batched together. States longer than 2,048 tokens are truncated; change the limit with load(max_state_tokens=...). The model was trained on states up to 1,024 tokens.

Server

bev-decider serve --port 8008
curl -s localhost:8008/v1/systemone -H 'content-type: application/json' -d '{
  "state": "Order 1182 arrived with a cracked screen.",
  "questions": {
    "damaged": {"type": "noul", "instructions": "Was the item damaged on arrival?"}
  }
}'

POST /v1/systemone accepts {"state", "questions", "model"?} and returns {"model", "answers", "latency_ms"}. GET /v1/models describes the loaded model. The server has no authentication and binds to 127.0.0.1 by default.

Command line

echo '{"state": "...", "questions": {...}}' | bev-decider decide

Tests

uv run pytest

The tests check that the package reproduces the probabilities of the released checkpoint on 50 fixed questions, in fp32 on CPU. They also check that answers do not depend on option order and that the server works.

License

The code is Apache-2.0. The model weights are CC-BY-NC-4.0 (see the model card).